The surprising part of automating legal work

We started with what seemed like a fairly ordinary exercise.

Take one area of business law — contract law — and identify all the work an advisory firm actually performs.

Not the lawyers.

Not the partners, associates, paralegals or business developers.

Just the work.

We mapped the lifecycle from identifying a potential assignment through legal production, managing the resulting contract and eventually identifying the next legal need.

The result was 45 distinct tasks.

Then we classified each one according to how far it could theoretically be moved from human execution to AI, agents and orchestration.

The result surprised me.

Of the 45 tasks:

19 appear capable of becoming fully orchestrated normal-flow tasks.

In other words, once the system has the necessary information, permissions, rules and sufficient demonstrated reliability, a human does not necessarily need to perform the task during normal execution.

Examples include things such as monitoring signals, collecting and structuring documents, retrieving information, monitoring obligations and deadlines, detecting deviations and identifying patterns across previous work.

Another 19 tasks appear capable of being performed substantially by agents, but with a human gate around material judgment or action.

The machine can do the work.

The human needs to review, approve, decide or take responsibility at the relevant point.

Only 7 of the 45 tasks remained in the category where the human should clearly continue to own the task itself, even if AI can do much of the preparation.

And even there, the reason was rarely that humans are uniquely good at searching documents, comparing clauses or producing first drafts.

The reasons were much more interesting:

intent, judgment, relationships, trade-offs, negotiation, decision and accountability.

That distinction matters.

Performing work is not the same as holding authority

Much of today's legal operating model implicitly combines two ideas:

The lawyer is responsible for the work.

Therefore:

The lawyer should perform the work.

AI makes that assumption increasingly questionable.

Responsibility and execution can be separated.

A machine may collect the evidence, reconstruct the chronology, identify the relevant questions, retrieve the applicable sources, produce an analysis, challenge that analysis and prepare the possible courses of action.

A human can then be brought in for the part that actually requires human authority:

What are we trying to achieve?

Which risk are we willing to take?

Which alternative should we choose?

Are we willing to stand behind this?

That is a very different role from manually performing every step leading up to the decision.

This does not mean removing humans

There is an important caveat.

The numbers above describe theoretical automation potential, not the governance model we should deploy on day one.

We are designing the platform around:

Human → Agent → Agent → Human.

A human gives the mandate.

One agent performs the work.

Another independently verifies or challenges it.

A human retains the required authority.

Over time, narrow and demonstrably reliable tasks may require less human intervention. But autonomy should be earned through evidence, not assumed because a model appears capable.

That may eventually mean that some of the 19 fully orchestratable tasks genuinely run without routine human involvement.

Others may always retain a human gate because the consequence of the decision demands it.

The interesting number may therefore be 38, not 19

The most important result of the exercise is perhaps not that 19 of 45 tasks could theoretically run without normal human intervention.

It is that 38 of the 45 tasks appear capable of being performed wholly or predominantly by machines, with humans entering primarily at defined gates.

That changes the architecture of legal work.

Instead of:

human performs process → technology assists

we can begin to imagine:

system performs process → human enters where judgment or authority creates value.

That is a much more profound change than adding a better legal chatbot.

It also changes what we should build

Once we saw this, it became obvious that building one AI tool for contract review, another for employment law and another for M&A would miss the point.

The underlying work repeatedly uses the same capabilities:

evidence, facts, research, reasoning, counter-analysis, decisions, drafting, monitoring and learning.

So we are designing one shared Legal Core and allowing different areas of law to operate as configurations over it.

Contract law is simply the first place we used to see the system clearly.

The larger question is now:

What does a full legal operating system look like when human attention becomes a deliberately allocated resource rather than the default execution engine for every task?

That is the problem we are working on now.